Most of these points are quite thoroughly described in the
/usr/src/linux/Documentation/networking/bonding.txt documentation file from the linux source package of your favorite distro. Speed of failover is controlled by the "miimon" parameter for most modes, but shouldn't be set too low; normal values are under one second anyway.
Here are the best parts, completed by me:
balance-rr or 0
Round-robin policy: Transmit packets in sequential
order from the first available slave through the
last. This mode provides load balancing and fault
active-backup or 1
Active-backup policy: Only one slave in the bond is
active. A different slave becomes active if, and only
if, the active slave fails. The bond's MAC address is
externally visible on only one port (network adapter)
to avoid confusing the switch.
This mode provides fault tolerance. The "primary"
option affects the behavior of this mode.
balance-xor or 2
XOR policy: Transmit based on the selected transmit
hash policy. The default policy is a simple [(source
MAC address XOR'd with destination MAC address) modulo
slave count]. Alternate transmit policies may be
selected via the xmit_hash_policy option.
This mode provides load balancing and fault tolerance.
broadcast or 3
Broadcast policy: transmits everything on all slave
interfaces. This mode provides fault tolerance.
802.3ad or 4
IEEE 802.3ad Dynamic link aggregation. Creates
aggregation groups that share the same speed and
duplex settings. Utilizes all slaves in the active
aggregator according to the 802.3ad specification.
Slave selection for outgoing traffic is done according
to the transmit hash policy, which may be changed from
the default simple XOR policy via the xmit_hash_policy
option. Note that not all transmit policies may be 802.3ad
compliant, particularly inregards to the packet mis-ordering
requirements of section 43.2.4 of the 802.3ad standard.
Differing peer implementations will have varying tolerances for
Note: Most switches will require some type of configuration
to enable 802.3ad mode.
balance-tlb or 5
Adaptive transmit load balancing: channel bonding that
does not require any special switch support. The
outgoing traffic is distributed according to the
current load (computed relative to the speed) on each
slave. Incoming traffic is received by the current
slave. If the receiving slave fails, another slave
takes over the MAC address of the failed receiving
balance-alb or 6
Adaptive load balancing: includes balance-tlb plus
receive load balancing (rlb) for IPV4 traffic, and
does not require any special switch support.
When a link is reconnected or a new slave joins the
bond the receive traffic is redistributed among all
active slaves in the bond by initiating ARP Replies
with the selected MAC address to each of the
clients. The updelay parameter must
be set to a value equal or greater than the switch's
forwarding delay so that the ARP Replies sent to the
peers will not be blocked by the switch.
balance-rr, active-backup, balance-tlb and balance-alb don't need switch support.
balance-rr augments performance at the price of fragmentation, performs poorly with some protocols (CIFS) and with more than 2 interfaces.
balance-alb and balance-tlb may not work properly with all switches; there are often some arp problems (some machines may fail to connect to each other for instance). You may need to tweak various settings (miimon, updelay) to get stable networking.
balance-xor may or may not require switch configuration. You need to set up an interface group (not LACP) on HP and Cisco switches, but apparently it's not necessary on D-Link, Netgear and Fujitsu switches.
802.3ad absolutely requires an LACP group on the switch side. It's the best supported option overall for augmenting performance.
Note: whatever you do, one network connection always go through one and only one physical link. So when aggregating GigE interfaces, a file transfer from machine A to machine B can't top 1 gigabit/s, even if each machine has 4 aggregated GigE interfaces (whatever the bonding mode in use).